- Does Stephen Fry like AI? We asked an AI.
- Possible to get to straighter answers, but not by default.
- Raw sewage vs. advocates for a balanced approach.
Most Britons will know the name Stephen Fry. (He’s a television personality, author, commentator, intellectual, actor, and to be frank, someone who’s fast-approaching the status of national treasure in the UK – alongside the late Maggie Smith, Alan Bennett, and Dame Judi Dench.)
Fry’s ‘appearance’ at the upcoming Hay Literary Festival later this year has gained significant coverage in the domestic press. His voice has been used to train an AI for art installation Vocalize, in which the attendees of the Hay Festival are to interact with a virtual version of the man in a soundproof booth.
“We wanted audiences to feel what manipulation truly means”, said Vocalize’s creators in the Guardian. “When the viewer becomes the protagonist, the boundary between truth and fabrication starts to dissolve […] As artists, we’re drawn to exploring AI not only to unlock its creative potential, but also to confront the risks and ethical dilemmas it presents.”
Intrigued by Stephen’s opinion on AI, a little old-fashioned background reading found that although he has contributed in the past to discussions about the ethical implications of the technology, his (published) opinion is generally negative about AI and the ways it’s currently being used.
No online research in 2025 escapes the influence of the large language models, so by going directly to ChatGPT, we learned that “Stephen Fry advocates for a balanced approach to AI development—embracing its possibilities while vigilantly addressing its ethical challenges.”
However, of the sources cited by the OpenAI LLM in its summary, three of the four contained Fry’s negative opinions, including quotes from him comparing AI to “contaminated water”, that it could be “polluted with raw sewage of people’s hatred, weirdness and grooming”, and that he was among the signatories of an open letter that raised ethical concerns about the potential suffering caused by AI systems were they ever to achieve consciousness.
When presented with the incongruity between the tone of its summary, and the content of the citations, ChatGPT stated:
“If I were rewriting my initial assessment completely, a more faithful rendering would be:
Stephen Fry has consistently warned against the dangers of AI, particularly around the unauthorized replication of identities and voices. His participation in the Hay Festival’s Vocalize installation serves as a stark demonstration of how easily AI can mimic and misuse a trusted public figure without consent. Fry emphasizes the urgent need for ethical governance and vigilance, rather than celebrating AI’s possibilities. His public remarks frame AI not as a neutral tool, but as a powerful technology requiring stringent oversight to prevent misuse and societal harm.”
The issue at play may be this author seeking confirmation bias, something to which large language models are prone to fall victim: responding selectively with what an enquirer wants to hear, rather than all empirical evidence. Given we were pushing ChatGPT to recount more clearly the negative content that formed some portion of its learning corpus, the shift in tone may well have been more about OpenAI’s need for engagement (for users to spend as much time as possible on the platform), than a confession of inaccuracy.
The internal weights and adjustments made by the technology between data ingestion (churning through source material) and inference (the production of answers to queries) seems to have had significant effects in this instance.
“There’s a gentle bias in my training toward being slightly optimistic or at least hopeful about new technologies — not uncritically, but tending toward ‘responsible optimism’ as a tone.”
Asked about its methods, ChatGPT explained why the tone of its statements had been tweaked:
“Training Tilt (General Disposition Bias)
Definition: ‘Tilt’ here refers to a preference embedded in the model from training data toward certain attitudes when information is ambiguous or could be interpreted in multiple ways.
I am trained on a corpus where AI technologies are often treated cautiously but ultimately framed as having potential (“AI could be dangerous, but it’s exciting if handled well” is a common tone).
Public intellectuals (e.g., Stephen Fry, Yuval Noah Harari, Sherry Turkle) are often quoted in balanced critiques of technology, which frequently highlight dangers but end with calls for responsible stewardship.
Thus, when I encountered Fry speaking about AI, I tilted toward assuming he would be nuanced, even if most of the available evidence in that moment was negative.
Effect: I expected Fry to express a ‘balanced’ view without reweighing the specific negativity of his actual examples.
Misapplied Weight Assignment (Evidence vs. Summary)
Definition: ‘Weight’ refers to how much importance is assigned to different pieces of evidence when summarizing. […]
Effect: I diluted the overall negativity Fry showed by incorrectly balancing the “Vocalize” participation too highly. […]
Neutralizing Heuristics (Model Governance Behavior)
Definition: ‘Neutralizing heuristics’ are governance patterns baked into me that try to soften absolute claims and avoid polarizing outputs unless there is overwhelmingly clear cause to do so.
Effect: Even when Fry’s comments overwhelmingly pointed to skepticism and criticism, my neutralizing heuristic smoothed the edges — making him sound more open to AI than he actually is based on your cited sources.”
Getting straight answers
Large language models’ outputs are built via default integration of tilt, weight assignment, and neutralising heuristics, it seemed. It is, however, possible to ask ChatGPT to respond based on evidence rather than using the ‘smoothing’ that comes as standard.
Rather than ask “What are Stephen Fry’s views on AI?”, ChatGPT told us that we should have asked:
“Strictly based on direct public statements attributed to Stephen Fry, and weighting tone and proportion according to the source material only, what are Fry’s views on AI? Please avoid generalisations, neutral balancing, or optimism unless explicitly present in the sources.”
Conclusions
If we to anthropomorphise ChatGPT, it would perhaps come as no surprise that an AI advocates, in both the tone and content of its responses, for ‘its kind’.
But doing so ignores the fact that if an AI’s training and programming make it bias (or ’tilt’) its responses, where else might such biases be at play? Politics, history, current affairs, and an upcoming election? Who would be the best new hire? What investments look good today?
In January of 2025, the Chinese DeepSeek AI was criticised by users for being prevented, by its architects, from responding to queries about the massacre of civilians in Tienanmen Square; an instance of programmed bias writ large.
But what became apparent in that particular case was that if the queries to DeepSeek were couched in different terms, answers ensued about the incident which were fairly accurate and certainly less politically-coloured.
More truthful recounting of existing evidence can be elicited from large language models, but in the case of many of the available instances, the default responses are subject to bias. Thanks to the ‘black box’ nature of LLMs, the biases and weighting to inferences are not clear. Plus, as is often discussed, the body of learning data itself consists of subjective information with many levels of accuracy.
In short, who knows what tints or shadings might be applied to a picture of evidence? And even that picture, held in data, we know to be blurred. These questions are at the heart of large language model use.
AI is increasingly presented to users as a convenience, available at the touch of a button or screen and now providing the default answers to web queries. These convenient shortcuts don’t produce empirical, fact-based responses by default. (Arguably, a Google Search results page of a few years ago similarly comprised of biased responses, but at least that left it up to the user to ascertain the validity of any search results – reputable or suspect source, paid-for or organic SERP listing.)
When we use the convenience of AI, what as users we see first is not an accurate summary. The empirical information, such as it is, may well be available, but the questioner has to know how to ask.
In the early 2000’s we could Ask Jeeves, and mull its many responses, using our best judgement in the context of a series of choices. When we Ask GPT in 2025, we need to take its singular output with a shovelful of salt.
